Applied AI in Instrumentation
and Measurement: The Deep
Learning Revolution
Mounib Khanafer and Shervin Shirmohammadi
I
n the last few years, hardly a day goes by that we do not
hear about the latest advancements and improvements
that Artificial Intelligence (AI) has brought to a wide
spectrum of domains: from technology and medicine to sci-
ence and sociology, and many others. AI is one of the core
enabling components of the fourth industrial revolution that
we are currently witnessing, and the applications of AI are
truly transforming our world and impacting all facets of so-
ciety, economy, living, working, and technology. The field of
Instrumentation and Measurement (I&M) is no exception,
and has already been impacted by Applied AI. In this article,
we give an overview of Applied AI and its usage in I&M. We
then take a deeper look at the I&M applications of one specific
AI method: Deep Learning (DL), which has recently revolu-
tionized the field of AI. Our survey of DL papers published
in the IEEE Transactions on Instrumentation and Measurement
(IEEE TIM) and IEEE Instrumentation & Measurement Magazine
showed that, since 2017, there is a very strong interest in apply-
ing DL methods to I&M, in terms of measurement, calibration,
and other I&M challenges. In particular, of the 32 surveyed pa-
pers, 75% were published in 2017 or later, and a remarkable
50% were published in 2019 alone. Considering that 2019 was
not yet finished when we were writing this article, the recent
exponential interest in and impact of DL in I&M is a very evi-
dent trend. We also found that although DL is used in a variety
of I&M topics, a considerable portion of DL in I&M focuses on
Vision Based Measurement (VBM) systems (around 28%) and
fault/defect diagnosis/detection/prediction (around 25%).
Finally, we found that Convolutional Neural Networks are the
most widely used DL technique in I&M, especially in VBM.
But to explain all of the above findings, we first need to under-
stand AI itself and what we mean by it in its applied context. So
let us begin our discussion with Applied AI.
Applied Artificial Intelligence
Although the long-term goal of research in AI is to enable
machines to have the same level of intelligence as animals or hu-
mans, it is important to note that the Applied AI of today is not
really comparable to biological intelligence. In fact, the word
10 IEEE Instrumentation & Measurement Magazine September 2020
1094-6969/20/$25.00©2020IEEE
Authorized licensed use limited to: Judy Scharmann. Downloaded on January 22,2021 at 15:02:28 UTC from IEEE Xplore. Restrictions apply.
Intelligence in the Applied AI of today is misleading for the com-
mon person, as it gives the person the wrong impression that he
or she is dealing with an intelligent being manifested as AI. This
becomes even more misleading if we consider that we do not
even have a universally agreed-upon definition of intelligence!
In other words, experts do not yet completely understand or
agree what intelligence is. So, then, how can we call something
artificial intelligence if we do not even know what intelligence is?
This has been an ongoing debate among AI experts for a long
time. For brevity, in this article we do not enter this debate, and
we refer the interested readers to other sources such as [1].
So, what does it mean when we talk about Applied AI in
today’s context, if it does not mean biological intelligence?
Today, Applied AI practically means the application of AI
methods as tools to advance or improve a system in a given do-
main; for example using AI methods to improve the weather
forecasting system in meteorology, increase the efficiency of
warehouse logistics in the storage and shipping industry, ob-
tain earlier diagnosis of diseases in medicine, or in the I&M
domain, reduce complexity of a measurement method or in-
crease accuracy of a measurement instrument. Researchers
expect that, in the future, these AI methods will match or even
surpass human intelligence. But today we are neither close
to that goal, nor do we need to be close in the applied arena,
because AI methods as they are today are already having a
beneficial impact on existing systems.
So, what are these methods? Fig. 1 shows some of AI’s most
commonly applied methods, which are normally inspired by
how the biological brain works, referred to as computational
intelligence [2]; for example, artificial neural networks try to
emulate the neural networks in a biological brain, or fuzzy logic
tries to operate similar to how humans make decisions without
complicated mathematical modeling and using only impre-
cise or vague information. What has made Applied AI such a
disruptive technology is the fact that these methods allow us
to do tasks such as classification, clustering, prediction, deci-
sion making, and optimization without the need to first build
an analytical model of the problem or the system at hand. This
is very important, so let us take a closer look at this.
 contributions to “meth-
ods or instruments for
measurement, detection,
tracking, monitoring, char-
acterization, identification,
estimation, or diagnosis of
a physical phenomenon;
uniqueness of an applica-
tion furthering the I&M
fields; or measurement
theory including uncer-
tainty, calibration, etc.” The
popularity of AI in I&M is
therefore not surprising,
considering that AI meth-
ods are quite suitable for
detection, tracking, mon-
itoring, characterization,
identification, estimation,
Fig. 1. Some of the most-commonly used Applied AI methods. Deep Learning, the subject of this article, is indicated in diagnosis, or prediction.
orange. Note that the methods shown in the figure are not exhaustive. In fact, the huge potential
of using AI in I&M was al-
Most of these methods take as input a set of previously gath- ready noticed by the 1990s, for example in [3], which describes
ered data and try to come up with a model that best matches this very well neural networks, fuzzy logic, and expert systems,
input data to the desired output. This model can then be used plus their applications in I&M, such as sensor design, calibra-
to obtain output values for new incoming data. This means that tion, uncertainty prediction, measurement data interpolation,
the model that an AI method creates is purely based on match- software instruments, indirect measurement, fault detection,
ing input to output (black box), unlike the analytical model that and system identification. AI is especially practical when the
a domain expert creates (white box). For example, let us say precise and accurate mathematical modeling of the measure-
we want to create a system that detects skin cancer from the ment system or instrument at hand is highly complex, highly
image of a patient’s skin lesions. To create such a system ana- nonlinear, highly dynamic, or impossible due to lack of knowl-
lytically, a biomedical engineer needs to develop an analytical edge of the system except for a limited number of parameters.
model based on the skin lesions’ shape, pattern, brightness, co- In such cases, or when the accuracy of the final measurement
lour, concentration, area, and many other parameters, some of is more important than understanding exactly how the sys-
which are complicated to model, impossible to model, or sim- tem works, AI can offer an attractive and practical solution.
ply unknown. Development of such an analytical model could One example that well illustrates the above notion is Vision-
be very complex, especially if the model needs to be generaliz- Based Measurement (VBM) [4], which requires AI as a core
able; i.e., working for any random skin type or color. There are component. An example of a VBM system is [5], which indi-
too many parameters here to come up with a general model rectly measures the amount of calories and nutrition in food
that would work with acceptable accuracy. But, if we use an AI from the food’s picture. Such a system would simply not be
method, for example a machine learning algorithm, we only implementable without AI, due to the highly complex nature
need to train it with a large-enough dataset of previously taken of the problem and lack of complete knowledge of all param-
images of skin lesions with and without cancer. The method eters. Another area where AI can be applied is calibration, for
then creates its own model of how to match features in those example in [6], where a picture of the user’s hand is used to
images to whether or not cancer is present. This not only signif- easily calibrate a complex force measuring instrument. Mea-
icantly reduces complexity, but also in some cases gives an even surement prediction is another useful feature of AI, which can
better result than an analytical model. Of course, the accuracy be applied when actual measurement is either costly or im-
of this AI-based model depends on the specific AI method and practical. For example, in a massively multisource networked
algorithm selected, as well as the quality of the provided data- system, it is not possible to explicitly measure the end-to-end
set. With this explanation of Applied AI in mind, let us now take delay between all pairs of nodes on the network, due to the
a look at how we can use it in I&M. O(N2) complexity of the problem. In such a case, AI has been
shown to predict measurements more accurately and orders of
Applied AI in I&M magnitude faster than non-AI techniques [7].
Looking at the scope webpage of IEEE Transactions on Instru- In Fig. 1 we can also see Deep Learning (DL), the main sub-
mentation & Measurement (IEEE TIM), (http://tim.ieee-ims. ject of this article. Although DL has existed since the 1980s [8],
org), we can say that I&M as a research field is interested in it wasn’t until 2011 and 2012 that it caused significant publicity

September 2020 IEEE Instrumentation & Measurement Magazine 11

Authorized licensed use limited to: Judy Scharmann. Downloaded on January 22,2021 at 15:02:28 UTC from IEEE Xplore. Restrictions apply.
among researchers and practitioners, when DL-based visual that GPUs can assist tremendously. Finally, the output layer
object recognition systems such as IDSIA and AlexNet beat produces the final output of the neural network.
out their competition by huge margins in terms of improved How a DL algorithm learns about features and manages
error rate [9], thanks partly to GPUs providing the significant to extract them depends on the availability of the training da-
speed up needed by DL. Another reason for their popularity is taset. Like any machine learning technique, the DL approach
that DL algorithms automate the process of extracting discrim- of learning can be done as either supervised learning or un-
inative features in a dataset, which usually requires domain supervised learning. In supervised learning, the dataset is
knowledge and significant human efforts [10]. This is why DL comprised of labeled data. That is, we deal with input data for
has revolutionized Applied AI in recent years. To understand which the respective output is known and defined. In unsu-
how DL can achieve such great results, and how it is being pervised learning, however, we deal with unstructured data
used in I&M, let us now take a look at the basics of DL. for which the output is unknown.
Training the DL network with the training dataset works as
Deep Learning Basics follows. As data flows from the input layer to the first hidden
A basic understanding of neural networks, as described in [3], layer, the neurons in this layer use their activation functions to
can be helpful to the readers here. A DL algorithm is based
on a layered architecture of data representation in which the
high-level features are extracted from the last layers of the neu-
ral network while the low-level features are extracted from the
lower layers [11]. The true inspiration behind such an archi-
tecture is to mimic how the biological brain works—as brains
extract data representation from an input, e.g., scene informa-
tion from the eyes, the generated output is a classified object [11].
One main capability of DL is to extract complex features
from a huge amount of data and discover hidden patterns and
trends in them. This is achieved by the utilization of a neural
network that is constituted by a set of interconnected neurons
(computational or processing units). A single neuron receives
data from inputs or other neurons, multiplies them with
weights and then feeds them through an activation function to
produce an output as depicted in Fig. 2.
In a neural network, neurons are organized into an input Fig. 2. A depiction of the operation of a neuron: it receives inputs (like x1 and
layer, one or more hidden layers, and an output layer as shown x2), multiplies them with weights (like w1 and w2), sums up the multiplication
results, and finally passes the sum to an activation function that produces the
in Fig. 3. The input layer is composed of input neurons that feed output f (x,w).
in the input data. A hid-
den layer receives its input
data either from the input
layer neurons or the neu-
rons of a preceding hidden
layer. Hidden layers per-
form intensive processing
of the data that were origi-
nally supplied by the input
layer. Apparently, the more
hidden layers the neural
network has, the deeper and
more intense processing
the data go through. In case
the neural network archi-
tecture is designed with
two or more hidden layers,
we end up with a DL archi-
tecture. Definitely, the use
of more hidden layers en-
tails the need for higher
computational and pro-
cessing capability, an area Fig. 3. Architecture of a neural network. The W1, W2, ... , W14 parameters are the weights of the connections.

12 IEEE Instrumentation & Measurement Magazine September 2020

Authorized licensed use limited to: Judy Scharmann. Downloaded on January 22,2021 at 15:02:28 UTC from IEEE Xplore. Restrictions apply.
produce outputs, as explained earlier, that are fed into the sec- estimation or Ensemble Averaging), Dropout Ensembles,
ond hidden layer. The latter repeats what the previous hidden Quantile Regression, and Gaussian Process Inference. Details
layer performed and feeds its outputs as inputs into the next are beyond the scope of this article, and in fact we are writing
hidden layer. This process is repeated until the output layer an article just about uncertainty in DL-based measurements
produces the final output. If the final output differs from the which we hope will be published in this same magazine in the
foreseen output, the weights on the connections between the near future. With the above general description about DL and
layers are adjusted and the whole process is repeated starting its uncertainty in mind, let us now take a look at specific DL
from the input layer. In each iteration of training, the amount techniques used in practice.
of deviation between the generated output and the expected
output is measured; this is referred to as the loss function. The Deep Learning Techniques
ultimate goal is to bring the loss function as close to zero as The DL techniques that we observed in our literature search
possible, making the DL algorithm more accurate in extracting include recurrent neural networks, convolutional neural net-
the correct features or patterns from a given set of input data. works, deep Boltzmann machines, deep belief networks,
But this also means that DL inherently has some level of uncer- generative adversarial networks, and autoencoders.
tainty, which we discuss next.
Recurrent Neural Networks
Uncertainty in Deep Learning Recurrent neural networks (RNNs) are developed to process
DL applications vary in the level of uncertainty they can tol- sequential information. In these networks, the same task is ap-
erate. While we can tolerate most Siri mistakes that appear plied for every element in the sequence of information, with
regularly, critical applications like autonomous driving or the output being dependent on the previous computations
content filtering for copyright legislation require models with [14]; thus, the name recurrent. The typical architecture of an
much less uncertainty in their predictions. Therefore, let us RNN is illustrated in Fig. 4a. In this figure, N designates part of
shed some light on the uncertainty in DL techniques and how a neural network, Xt refers to inputs over time, and Yt refers to
it is addressed by researchers. the corresponding outputs over time.
Like any other measurement system, two types of effects It is quite convenient to unfold the architecture in Fig. 4a to
can contribute to uncertainty in a DL system, namely: sys- look like the one in Fig. 4b. In this figure, we can clearly see that
tematic and random [12]. Systematic effects occur when the a type of memory is employed in RNNs where the information
DL system is trained with a dataset that is not sufficiently rep- that has been calculated till time t is preserved. One popu-
resentative of the entire input domain, or if the DL model lar model for an RNN is the long short-term memory (LSTM)
underfits the training data. Even if a DL model does not un- model that was proposed by Gers and Schmidhuber in [15]. Typ-
derfit the training data, it can reasonably predict measurement ical applications that benefit from RNNs are speech recognition,
results based on data patterns that it has been trained with be- language modeling, translation, and image captioning [14].
fore; uncertainty may rise when it is fed with data patterns that
it is seeing for the first time. This can be improved by training Convolutional Neural Networks
the DL with more generalizable data, and with using better hy- Convolutional neural networks (CNNs) employ convolu-
pothesis functions. Random effects, on the other hand, arise if tional layers, subsampling (pooling) layers, and a final stage of
during prediction (not during training) the DL model slightly a fully connected layer, as illustrated in Fig. 5. These networks
changes, such as with Monte Carlo Dropout, which causes the are named convolutional due to their use of the convolution
DL model to produce different outputs for the same input de- mathematical operation in their feature extraction process.
pending on the nodes that
are dropped. These ran-
dom effects will be larger
if the DL model suffers
from overfitting during
training. Whether we are
dealing with systematic or
random effects, it is essen-
tial to quantify the level
of uncertainty associated
with the DL techniques. In
DL, we can typically use
the following methods to
quantify the uncertainty
[13]: Monte-Carlo Drop-
out, Deep Ensembles (e.g.,
Distributional Parameter Fig. 4. a) The typical architecture of an RNN. b) Unfolded RNN architecture.

September 2020 IEEE Instrumentation & Measurement Magazine 13

Authorized licensed use limited to: Judy Scharmann. Downloaded on January 22,2021 at 15:02:28 UTC from IEEE Xplore. Restrictions apply.
 nodes form full connectiv-
ity across subsequent layers.
The visible nodes are the ones
that receive the training set of
data. Compared to other DL
networks, DBMs are unique
in that the input visible nodes
are connected to each other.
The ultimate goal of a DBM
is to detect the distribution of
the given input training da-
taset. Typical applications of
DBMs are speech and object
Fig. 5. Typical architecture of a CNN. recognition [11].
Deep belief networks
Both convolution and subsampling layers are used for fea- (DBNs) are based on stacking RBM layers on top of each other.
ture extraction. Specifically, as shown in Fig. 5, convolution However, while top layers have undirected edges, lower lay-
layers produce feature maps while subsampling layers reduce ers have directed ones, as shown in Fig. 6b.
the sizes of these maps without losing the key information in
them. The final output of these layers is fed into the fully con- Generative Adversarial Networks
nected layer which handles the task of classification. CNNs are Generative adversarial networks (GANs) suggest an approach
typically used in computer vision and speech recognition ap- for maximum likelihood estimation and employ two neural net-
plications [11]. works that compete against each other in a zero-sum game [16].
Basically, the GAN architecture utilizes a generator model and a
Deep Boltzmann Machines and Deep Belief discriminator model. The generator is the system that is trained
Networks to generate images, while the discriminator is the system that
In a Boltzmann machine, nodes are fully connected to each classifies these images as accurate or not. This process of image
other using undirected edges, as shown in Fig. 6a. The nodes generation and classification is repeated until the generator is able
are divided into visible and hidden nodes, with no output to produce accurate results. Examples of applications of GANs
layer of nodes. In restricted Boltzmann machines (RBMs), include game development and artificial video generation [16].
however, the visible nodes are only connected to hidden
nodes, and vice versa. Deep Boltzmann machines (DBMs) are Autoencoder
formed by stacking RBMs on top of each other. The hidden Autoencoding is a DL algorithm that efficiently com-
presses information and
learns how to reproduce ac-
curate approximation of the
original information from
the compressed data. Since
they are used for data com-
pression, autoencoders are
efficient tools for dimension-
ality reduction. Typical where
autoencoders are applied in-
clude data denoising and
dimensionality reduction for
data visualization [14].

Deep Learning in
Instrumentation
and Measurement
Here, we describe examples
of how DL is being used in
I&M literature, with a note
Fig. 6. a) Deep Boltzmann machine (DBM). The black nodes are hidden neurons while the white nodes are visible
that our search was inex-
neurons. b) Architecture of the deep belief network (DBN). The black nodes are hidden neurons and organized into
several layers while the white nodes are visible neurons. Lower layers use directed edges as opposed to fully haustive and mostly limited
undirected edges in DBMs. to DL papers appearing in

14 IEEE Instrumentation & Measurement Magazine September 2020

Authorized licensed use limited to: Judy Scharmann. Downloaded on January 22,2021 at 15:02:28 UTC from IEEE Xplore. Restrictions apply.
IEEE TIM. We categorize the examples based on the DL tech-
nique used.
CNN
CNN has been used extensively in VBM systems. For example,
a three-stage automatic defect inspection system for high-
speed railways has been proposed in PVANET++ [17]. As split
pins have a key role in fixing joint components on catenary
support devices (CSDs), PVANET++ localizes and inspects
split pins by using CNN to detect defects. Another system uses
a two-stage deep CNN to detect insulator surface defects in
railway catenary [18] by firstly localizing the catenary compo-
nents to obtain images of the insulators, and then using a deep
material classifier and a deep denoising autoencoder to detect
defects. Since the defect detection of the fasteners on a CSD is
essential for both safety and cost reductions in the operation
of high-speed railways, a VBM method based on a deep CNN
[19] is used to tackle the design of such detection systems.
A VBM system for the purpose of measuring pain intensity
through the analysis of facial expressions is presented in [20]
which uses the aforementioned AlexNet architecture to extract
critical features from patients’ images and to draw conclusions
on the level of pain they are experiencing. The work in [21] pro-
poses a vision-based evaluation (VBE) framework for VBM
systems that studies the capability of DL strategies to deal with
uncertainty contributions, a very important issue in I&M. This
framework is also centered around an AlexNet-based CNN and
handles uncertainty contributions during calibration processes.
Another VBM system is presented in [22] which uses both CNN
and DBM to design a food recognition system that analyzes pic-
tures of meals, taken by a mobile device, to identify different food
items and estimate their calories and nutrition. Yet another VBM
system [23] tackles thermal image processing. A dataset of high-
resolution thermal facial images is built and used to train the deep
alignment network (DAN) algorithm (which is based on CNN)
for face analysis. CNN is also used in robotic VBM systems,
for example [24] uses a Faster Region CNN to aid robots in the
recognition of hand gesture, while [25] introduces a projection al-
gorithm to generate RGB, depth, and intensity (RGB-DI) images
to measure the outdoor environments with a variable resolution.
A full CNN is then used to segment those RGB-DI images and use
them to realize the 3-D scene understanding for mobile robots.
The application of DL principles in light detection and
ranging (LiDAR)-based perception tasks is also studied, for
example, in [26]. This study, due to the limited availability of
LiDAR point cloud datasets, uses simulators to automatically
generate 3-D annotated LiDAR point clouds, and then uses
that data to train a deep model that incorporates both CNN
and RNN principles. Since network binarization is associated
with reductions in computational and memory costs in 2-D
computer vision tasks, [27] proposes a technique to train bi-
nary volumetric CNNs for 3-D object recognition.
CNN has also been used for extracting features from vi-
bration signals to aid in the diagnosis of faults of a magnet
synchronous motor [28], fault diagnosis of spindle bearings
[29], and safety of the pipelines used for liquid petroleum
September 2020 IEEE Instrumentation & Measurement Magazine 15
Authorized licensed use limited to: Judy Scharmann. Downloaded on January 22,2021 at 15:02:28 UTC from IEEE Xplore. Restrictions apply.
transportation [30]. As magnetic flux leakage (MFL) is a
common testing method for those purposes, the defects are
identified from MFL images based on CNNs.
RNN
RNNs have been used in the detection of fundamental pha-
sors and the identification of control and protection signals in
power systems [31], as well as estimating lithium-ion battery
remaining useful life to achieve an intelligent battery manage-
ment system [32], and supporting predictive models for voltage
correction in a dc voltage reference source [33]. In nonlinear dy-
namical systems, RNNs are used to tackle the identification
problem: finding a time-dependent model for the behavior of
the process generating the data [34], as well as in the modelling
of dynamic systems in the absence of measurable state variables
[35]. Another use of RNNs is in modeling of processes that oc-
cur in various industries like steam-raising plants, gas turbines,
and automotive suspensions [36], since the output of such pro-
cesses has unsymmetrical behavior. Finally, RNNs have been
used for fault detection: in wireless sensor networks for sensor
node fault detection [37], health condition monitoring of a ma-
chine [38], or early fault detection in industrial applications [39],
where a deep neural network is used for feature extraction and
the LSTM network for distribution estimator.
DBN
A discriminative DBN and ant colony optimization model has
been built in [40] for the purpose of health condition monitor-
ing of machines. DBN is also used to distinguish acceptable
and unacceptable segments in an electrocardiogram (ECG)
signal [41], to reduce false alarms during atrial fibrillation
detection. Finally, DBN has been used in a soft sensor for esti-
mating the deflection of a polymeric mechanical actuator [42].
The latter is based on ionic polymer-metal composites, which
are used in important fields like robotics and surgery.
Autoencoders
Autoencoders are used in rotating machinery measurements,
for example in fault identification for rotating machines [43] us-
ing a stacked sparse autoencoders-based deep neural network
coupled with the concept of compressed sensing, or in studying
the conditions of rotating machinery [44], where sensors gener-
ate fault signals, and features are extracted from these signals
and fused by sparse autoencoder (SAE) neural networks. The
fused features are then classified by a DBN. Another work that
uses autoencoders for fault diagnosis is based on the analysis of
vibration signals in wind turbine gearboxes [45], where stacked
multilevel-denoising autoencoders are used to assist in learn-
ing discriminative representations of fault features.
Localization is another application of autoencoders. For
example, the human step length in Pedestrian Dead Reckon-
ing systems can be estimated with stacked autoencoders [46].
Finally, deep autoencoders are used for feature extraction
to mitigate the corruption of Photoplethysmographic (PPG)
measurements due to motion artifacts in personal healthcare
systems [47].
Generative Adversarial Network (GAN)
GAN is one of the newest members of the DL family, and so
it is not surprising not to see as many applications with GAN
as with the other DL techniques. Among the papers we stud-
ied, GAN was used to measure the reliability of transmission
gears [48].
Observations and Future Trends
The following observations can be made from the studied
literature:
◗◗ The spectrum of applications benefiting from DL is quite
broad. Researchers from various fields (industrial and
systems engineering, operations, food control, health
care, machinery, transportation, image processing, circuit
design, sensor networks, etc.) are benefiting from DL
architectures. This ought to drive DL research in I&M
forward and pave the way for promising solutions to
known and/or unresolved issues.
◗◗ Several papers stress that their proposed solutions are the
first to apply DL to their problem of interest. This popular-
ity of using DL is stimulated by the fact that an abundance
of datasets is becoming available for training DL algo-
rithms. In earlier years, many applications suffered from
the scarcity of the available training data.
◗◗ There is a strong focus on using CNN. The reason is that
CNNs are proven to be efficient in computer vision appli-
cations [12], so they are the natural first choice in VBM
systems. The fact that rich datasets of images and videos
have been recently made available for training purposes
boosts the usage of AI and of CNN in particular.
◗◗ A considerable percentage of the publications focus on
VBM (around 28%) and on fault/defect diagnosis/detec-
tion/prediction (about 25%). Many industry sectors can
highly benefit from such research.
Despite the above encouraging findings, two clear gaps
were observed in the limited literature that we searched:
◗◗ The usage of DL in biometric and security systems: DL can
be an excellent choice to design face recognition, identity
check, or other types of user identification components
for security systems.
◗◗ As Industry 4.0 unfolds, it appears as a major area to
leverage DL. The industrial Internet of Things (IoT) para-
digm in particular appears to be a strong candidate for the
utilization of DL architectures. Industrial IoT is heavily
involved in monitoring activities, calibration and control
of sensor nodes, fault detection, etc., which can bene-
fit from the capabilities of DL. Also, given the fact that a
huge amount of data is generated and processed within
an IoT platform, autoencoders can be an efficient tool to
compress data, which reduces the burden on the commu-
nication system.
References
[1] K. Warwick and S. J. Nasuto, “Historical and current machine
intelligence,” IEEE Instrum. Meas. Mag., vol. 9, no. 6, pp. 20-26,
Dec. 2006.
16 IEEE Instrumentation & Measurement Magazine September 2020
Authorized licensed use limited to: Judy Scharmann. Downloaded on January 22,2021 at 15:02:28 UTC from IEEE Xplore. Restrictions apply.
[2] D. B. Fogel, “Machine intelligence: teaching computers to learn,”
IEEE Instrum. Meas. Mag., vol. 9, no. 3, pp. 12-16, Jun. 2006.
[3] C. Alippi et al., “Artificial intelligence for instruments and
measurement applications,” IEEE Instrum. Meas. Mag., vol. 1, no.
2, pp. 9-17, Jun. 1998.
[4] S. Shirmohammadi and A. Ferrero, “Camera as the instrument:
the rising trend of vision based measurement,” IEEE Instrum.
Meas. Mag., vol. 17, no. 3, pp. 41-47, Jun. 2014.
[5] P. Pouladzadeh et al., “Measuring calorie and nutrition from food
image,” IEEE Trans. Instrum. Meas., vol. 63, no. 8, pp. 1947-1956,
Aug. 2014.
[6] J. Zhou et al., “A new hand measurement method to simplify
calibration in CyberGlove-based virtual rehabilitation,” IEEE
Trans. Instrum. Meas., vol. 59, no. 10, pp. 2496-2504, Oct. 2010.
[7] S. A. Mohammed et al., “A multimodal deep learning based
distributed network latency measurement system,” IEEE Trans.
Instrum. Meas., Jan. 2020.
[8] R. Dechter, “Learning while searching in constraint-satisfaction
problems,” University of California, in Proc. AAAI National Conf.
Artificial Intelligence, pp. 178-183, Aug. 1986.
[9] J. Schmidhuber, “History of computer vision contests won by
deep CNNs on GPU,” March 2017. [Online]. Available: http://
people.idsia.ch/∼juergen/computer-vision-contests-won-by-
gpu-cnns.html.
[10] M. M. Najafabadi et al., “Deep learning applications and challenges
in Big Data analytics,” J. Big Data, vol. 2, no. 1, pp. 1-21, Feb. 2015.
[11] S. Pouyanfar et al., “A survey on deep learning: algorithms,
techniques, and applications,” ACM Computing Surveys, vol. 51,
no. 5, Article 92, Sep. 2018.
[12] A. Ferrero and S. Salicone, “Measurement uncertainty,” IEEE
Instrum. Meas. Mag., vol. 9, no. 3, pp. 44-51, Jun. 2006.
[13] S. Bachstein, “Uncertainty quantification in deep learning,”
Master’s Thesis, Ulm University, Germany, Jan. 2019.
[14] M. Coşkun et al., “An overview of popular deep learning
methods,” European J. Technic., vol. 7, no. 2, pp. 165-176, Dec. 2017.
[15] F. A. Gers and J. Schmidhuber, “Recurrent nets that time and
count,” in Proc. IEEE-INNS-ENNS Int. Joint Conf. Neural Networks,
vol. 3, Jul. 2000.
[16] M. Z. Alom et al., “A state-of-the-art survey on deep learning
theory and architectures,” Electronics, vol. 8, no. 292, pp. 1-67,
Mar. 2019.
[17] J. Zhong et al., “A CNN-based defect inspection method for
catenary split pins in high-speed railway,” IEEE Trans. Instrum.
Meas., vol. 68, no. 8, pp. 2849-2860, Aug. 2019.
[18] G. Kang et al., “Deep architecture for high-speed railway insulator
surface defect detection: denoising autoencoder with learning,”
IEEE Trans. Instrum. Meas., vol. 68, no. 8, pp. 2679-2690, Aug. 2019.
[19] J. Chen et al., “Automatic defect detection of fasteners on the
catenary support device using deep convolutional neural network,”
IEEE Trans. Instrum. Meas., vol. 67, no. 2, pp. 257-269, Feb. 2018.
[20] P. Casti et al., “Calibration of vision-based measurement of pain
intensity with multiple expert observers,” IEEE Trans. Instrum.
Meas., vol. 68, no. 7, pp. 2442-2450, Jul. 2019.
[21] G. Callari et al., “The influence of uncertainty contributions on
deep learning architectures in vision-based evaluation systems,”
IEEE Trans. Instrum. Meas., vol. 68, no. 7, pp. 2425-2432, Jul. 2019.
[22] P. Pouladzadeh and S. Shirmohammadi, “Mobile multi-food
recognition using deep learning,” ACM Trans. Multimedia
Computing, Commun. Applications, vol. 13, no. 3s, article 36, Aug.
2017.
[23] M. Kopaczka et al., “A thermal infrared face database with facial
landmarks and emotion labels,” IEEE Trans. Instrum. Meas., vol.
68, no. 5, pp. 1389-1401, May 2019.
[24] C. Nuzzi et al., “Deep learning-based hand gesture recognition for
collaborative robots,” IEEE Instrum. Meas. Mag., vol. 22, no. 2, pp.
44-51, Apr. 2019.
[25] Z. Qiu et al., “RGB-DI images and full convolution neural
network-based outdoor scene understanding for mobile robots,”
IEEE Trans. Instrum. Meas., vol. 68, no. 1, pp. 27-37, Jan. 2019.
[26] F. Wang et al., “Automatic generation of synthetic LiDAR point
clouds for 3-D data analysis,” IEEE Trans. Instrum. Meas., vol. 68,
no. 7, pp. 2671-2673, Jul. 2019.
[27] C. Ma et al., “Binary volumetric convolutional neural networks
for 3-D object recognition,” IEEE Trans. Instrum. Meas., vol. 68, no.
1, Jan. 2019.
[28] I-H. Kao et al., “Analysis of permanent magnet synchronous
motor fault diagnosis based on learning,” IEEE Trans. Instrum.
Meas., vol. 68, no. 2, pp. 310-324, Feb. 2019.
[29] X. Ding and Q. He, “Energy-fluctuated multiscale feature learning
with deep ConvNet for intelligent spindle bearing fault diagnosis,”
IEEE Trans. Instrum. Meas., vol. 66, no. 8, pp. 1926-1935, Aug. 2017.
[30] J. Feng et al., “Injurious or noninjurious defect identification from
MFL images in pipeline inspection using convolutional neural
network,” IEEE Trans. Instrum. Meas., vol. 66, no. 7, pp. 1883-
1892Jul. 2017.
[31] I. Kamwa et al., “Recurrent neural networks for phasor detection
and adaptive identification in power system control and
protection,” IEEE Trans. Instrum. Meas., vol. 45, no. 2, pp. 657-663,
Jul. 1996.
[32] D. Liu et al., “An integrated probabilistic approach to lithium-ion
battery remaining useful life estimation,” IEEE Trans. Instrum.
Meas., vol. 64, no. 3, pp. 660-670, Mar. 2015.
[33] I. Nancovska et al., “Case study of the predictive models used for
stability improvement of the dc voltage reference source,” IEEE
Trans. Instrum. Meas., vol. 47, no. 6, pp. 1487-1491, Dec. 1998.
[34] C. Alippi and V. Piuri, “Experimental neural networks for
prediction and identification,” IEEE Trans. Instrum. Meas., vol. 45,
no. 2, pp. 670-676, Apr. 1996.
[35] C. Alippi and V. Piuri, “Neural modeling of dynamic systems
with nonmeasurable state variables,” IEEE Trans. Instrum. Meas.,
vol. 48, no. 6, pp. 1073-1080, Dec. 1999.
[36] A. H. Tan and K. Godfrey, “Modeling of direction-dependent
processes using wiener models and neural networks with
nonlinear output error structure,” IEEE Trans. Instrum. Meas., vol.
53, no. 3, pp. 744-753, Jun. 2004.
[37] A. I. Moustapha and R. R. Selmic, “Wireless sensor network
modeling using modified recurrent neural networks: application
to fault detection,” IEEE Trans. Instrum. Meas., vol. 57, no. 5, pp.
981-988, May 2008.
[38] A. Malhi et al., “Prognosis of defect propagation based on
recurrent neural networks,” IEEE Trans. Instrum. Meas., vol. 60,
no. 3, pp. 703-711, Mar. 2011.
September 2020 IEEE Instrumentation & Measurement Magazine 17
Authorized licensed use limited to: Judy Scharmann. Downloaded on January 22,2021 at 15:02:28 UTC from IEEE Xplore. Restrictions apply.
[39] W. Lu et al., “Early fault detection approach with deep architectures,”
IEEE Trans. Instrum. Meas., vol. 67, no. 7, pp. 1679-1689, Jul. 2018.
[40] M. Ma et al., “Discriminative deep belief networks with ant colony
optimization for health status assessment of machine,” IEEE
Trans. Instrum. Meas., vol. 66, no. 12, pp. 3115-3125, Dec. 2017.
[41] B. Taji et al., “False alarm reduction in atrial fibrillation detection
using deep belief networks,” IEEE Trans. Instrum. Meas., vol. 67,
no. 5, pp. 1124-1131May 2018.
[42] B. Andò et al., “Low-order nonlinear finite-impulse response soft
sensors for ionic electroactive actuators based on deep learning,”
IEEE Trans. Instrum. Meas., vol. 68, no. 5, pp. 1637-1646, May 2019.
[43] J. Sun et al., “Intelligent bearing fault diagnosis method
combining compressed data acquisition and deep learning,” IEEE
Trans. Instrum. Meas., vol. 67, no. 1, pp. 185-195, Jan. 2018.
[44] Z. Chen and W. Li, “Multisensor feature fusion for bearing fault
diagnosis using sparse autoencoder and deep belief network,”
IEEE Trans. Instrum. Meas., vol. 66, no. 7, pp. 1693-1702, Jul. 2017.
[45] G. Jiang et al., “Stacked multilevel-denoising autoencoders: a new
representation learning approach for wind turbine gearbox fault
diagnosis,” IEEE Trans. Instrum. Meas., vol. 66, no. 9, pp. 2391-
2402, Sep. 2017.
[46] F. Gu et al., “Accurate step length estimation for pedestrian dead
reckoning localization using stacked autoencoders,” IEEE Trans.
Instrum. Meas., vol. 68, no. 8, pp. 2705-2713, Aug. 2019.
[47] M. Singha Roy et al., “Improving photoplethysmographic
measurements under motion artifacts using artificial neural
network for personal healthcare,” IEEE Trans. Instrum. Meas., vol.
67, no. 12, pp. 2820-2829, Dec. 2018.
[48] J. Li et al., “CGAN-MBL for reliability assessment with
imbalanced transmission gear data,” IEEE Trans. Instrum. Meas.,
vol. 68, no. 9, pp. 3173-3183, Sep. 2019.
Mounib Khanafer is an Associate Professor of Electrical and
Computer Engineering at the Department of Engineering at
the American University of Kuwait, Kuwait. He conducts re-
search in the general areas of Wireless Sensor Networks and
Internet of Things and their applications in vehicular networks.
He has three years of industrial experience at Nortel Networks,
Canada and worked as a Postdoctoral Fellow at the School of
Electrical Engineering and Computer Science at the University
of Ottawa, Canada from 2012-2013. He received his B.Sc. degree
in electrical engineering from Kuwait University, Kuwait in
2002 and his M.Sc. and Ph.D. degrees in electrical and computer
engineering from the University of Ottawa, Canada in 2007 and
2012, respectively. He is a Senior Member of IEEE.
Shervin Shirmohammadi (M ’04, SM ’04, F ’17) is currently a Pro-
fessor with the School of Electrical Engineering and Computer
Science, University of Ottawa, Canada, where he is Director
of the Distributed and Collaborative Virtual Environment Re-
search Laboratory and his research focuses on multimedia
systems and networks, including measurement techniques and
applied AI for networking, video streaming, and health sys-
tems. He received his Ph.D. degree in electrical engineering from
University of Ottawa, Canada and currently serves as the Editor-
in-Chief of IEEE Transactions on Instrumentation and Measurement.